-
Notifications
You must be signed in to change notification settings - Fork 176
/
gen_bottom_up_proposals.py
193 lines (161 loc) · 7.2 KB
/
gen_bottom_up_proposals.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
import argparse
import os
import sys
import math
import numpy as np
import multiprocessing
from sklearn.metrics import confusion_matrix
import time
import pickle
import multiprocessing as mp
from ops.sequence_funcs import *
from ops.anet_db import ANetDB
from ops.thumos_db import THUMOSDB
from ops.detection_metrics import get_temporal_proposal_recall, name_proposal
from ops.sequence_funcs import temporal_nms
from ops.io import dump_window_list
parser = argparse.ArgumentParser()
parser.add_argument('score_files', type=str, nargs='+')
parser.add_argument("--anet_version", type=str, default='1.2', help='')
parser.add_argument("--dataset", type=str, default='activitynet', choices=['activitynet', 'thumos14'])
parser.add_argument("--cls_scores", type=str, default=None,
help='classification scores, if set to None, will use groundtruth labels')
parser.add_argument("--subset", type=str, default='validation', choices=['training', 'validation', 'testing'])
parser.add_argument("--iou_thresh", type=float, nargs='+', default=[0.5, 0.75, 0.95])
parser.add_argument("--score_weights", type=float, nargs='+', default=None, help='')
parser.add_argument("--write_proposals", type=str, default=None, help='')
parser.add_argument("--minimum_len", type=float, default=0, help='minimum length of a proposal, in second')
parser.add_argument("--reg_score_files", type=str, nargs='+', default=None)
parser.add_argument("--frame_path", type=str, default='/mnt/SSD/ActivityNet/anet_v1.2_extracted_340/')
args = parser.parse_args()
if args.dataset == 'activitynet':
db = ANetDB.get_db(args.anet_version)
db.try_load_file_path('/mnt/SSD/ActivityNet/anet_v1.2_extracted_340/')
elif args.dataset == 'thumos14':
db = THUMOSDB.get_db()
db.try_load_file_path('/mnt/SSD/THUMOS14/')
# rename subset test
if args.subset == 'testing':
args.subset = 'test'
else:
raise ValueError("unknown dataset {}".format(args.dataset))
video_list = db.get_subset_videos(args.subset)
video_list = [v for v in video_list if v.instances != []]
print("video list size: {}".format(len(video_list)))
# load scores
print('loading scores...')
score_list = []
for fname in args.score_files:
score_list.append(pickle.load(open(fname, 'rb')))
print('load {} piles of scores'.format(len(score_list)))
# load classification scores if specified
if args.cls_scores:
cls_scores = cPickle.load(open(args.cls_scores, 'rb'))
else:
cls_scores = None
print('done')
# load regression scores
if args.reg_score_files is not None:
print('loading regression scores')
reg_score_list = []
for fname in args.reg_score_files:
reg_score_list.append(cPickle.load(open(fname, 'rb')))
print('load {} piles of regression scores'.format(len(reg_score_list)))
else:
reg_score_list = None
# merge scores
print('merging scores')
score_dict = {}
for key in score_list[0].keys():
out_score = score_list[0][key].mean(axis=1) * (1.0 if args.score_weights is None else args.score_weights[0])
for i in range(1, len(score_list)):
add_score = score_list[i][key].mean(axis=1)
if add_score.shape[0] < out_score.shape[0]:
out_score = out_score[:add_score.shape[0], :]
elif add_score.shape[0] > out_score.shape[0]:
tick = add_score.shape[0] / float(out_score.shape[0])
indices = [int(x * tick) for x in range(out_score.shape[0])]
add_score = add_score[indices, :]
out_score += add_score * (1.0 if args.score_weights is None else args.score_weights[i])
score_dict[key] = out_score
print('done')
# merge regression scores
if reg_score_list is not None:
print('merging regression scores')
reg_score_dict = {}
for key in reg_score_list[0].keys():
out_score = reg_score_list[0][key].mean(axis=1)
for i in range(1, len(reg_score_list)):
add_score = reg_score_list[i][key].mean(axis=1)
if add_score.shape[0] < out_score.shape[0]:
out_score = out_score[:add_score.shape[0], :]
out_score += add_score
reg_score_dict[key] = out_score / len(reg_score_list)
print('done')
else:
reg_score_dict = None
# bottom-up generate proposals
print('generating proposals')
pr_dict = {}
pr_score_dict = {}
topk = 1
def gen_prop(v):
if (args.dataset == 'activitynet') or (args.dataset == 'thumos14'):
vid = v.id
else:
vid = v.path.split('/')[-1].split('.')[0]
scores = score_dict[vid]
frm_duration = len(scores)
topk_cls = [0]
topk_labels = label_frame_by_threshold(scores, topk_cls, bw=3, thresh=[0.01, 0.05, 0.1, .15, 0.25, .4, .5, .6, .7, .8, .9, .95, ], multicrop=False)
bboxes = []
tol_lst = [0.05, .1, .2, .3, .4, .5, .6, 0.8, 1.0]
bboxes.extend(build_box_by_search(topk_labels, np.array(tol_lst)))
if reg_score_dict:
reg_scores = reg_score_dict[score_id]
bboxes = regress_box(bboxes, reg_scores, len(scores))
# print len(bboxes)
bboxes = temporal_nms(bboxes, 0.9)
pr_box = [(x[0] / float(frm_duration) * v.duration, x[1] / float(frm_duration) * v.duration) for x in bboxes]
# filter out too short proposals
pr_box = list(filter(lambda b: b[1] - b[0] > args.minimum_len, pr_box))
return v.id, pr_box, [x[3] for x in bboxes]
def call_back(rst):
pr_dict[rst[0]] = rst[1]
pr_score_dict[rst[0]] = rst[2]
import sys
print(rst[0], len(pr_dict), len(rst[1]))
sys.stdout.flush()
pool = mp.Pool(processes = 32)
lst = []
handle = [pool.apply_async(gen_prop, args=(x, ), callback=call_back) for x in video_list]
pool.close()
pool.join()
# evaluate proposal info
proposal_list = [pr_dict[v.id] for v in video_list if v.id in pr_dict]
gt_spans_full = [[(x.num_label, x.time_span) for x in v.instances] for v in video_list if v.id in pr_dict]
gt_spans = [[item[1] for item in x] for x in gt_spans_full]
score_list = [score_dict[v.id] for v in video_list if v.id in pr_dict]
duration_list = [v.duration for v in video_list if v.id in pr_dict]
proposal_score_list = [pr_score_dict[v.id] for v in video_list if v.id in pr_dict]
print('{} groundtruth boxes from'.format(sum(map(len, gt_spans))))
print('average # of proposals: {}'.format(np.mean(list(map(len, proposal_list)))))
IOU_thresh = np.arange(0.5, 1, 0.2)
p_list = []
for th in IOU_thresh:
pv, pi = get_temporal_proposal_recall(proposal_list, gt_spans, th)
print('IOU threshold {}. per video recall: {:02f}, per instance recall: {:02f}'.format(th, pv * 100, pi * 100))
p_list.append((pv, pi))
print('Average Recall: {:.04f} {:.04f}'.format(*(np.mean(p_list, axis=0)*100)))
if args.write_proposals:
name_pattern = 'img_*.jpg'
frame_path = args.frame_path
named_proposal_list = [name_proposal(x, y) for x, y in zip(gt_spans_full, proposal_list)]
allow_empty = args.dataset == 'activitynet' and args.subset == 'testing'
dumped_list = [dump_window_list(v, prs, frame_path, name_pattern, score=score, allow_empty=allow_empty) for v, prs, score in
zip(filter(lambda x: x.id in pr_dict, video_list), named_proposal_list, score_list)]
with open(args.write_proposals, 'w') as of:
for i, e in enumerate(dumped_list):
of.write('# {}\n'.format(i + 1))
of.write(e)
print('list written. got {} videos'.format(len(dumped_list)))